MC Frequency Dropout for CNNs

• MC Frequency Dropout is a stochastic regularization method that suppresses selected frequency components to generate coherent uncertainty maps in CNNs.
• It integrates with convolutional architectures by applying FFT-based dropout after convolution blocks, preserving spatial integrity and anatomical details.
• Empirical results show improved segmentation accuracy (DSC) and reduced uncertainty calibration error, underscoring its robustness against imaging noise.

MC Frequency Dropout (Monte Carlo Frequency Dropout) is a stochastic regularization and Bayesian inference strategy for convolutional neural networks (CNNs), in which signal or feature map components are attenuated or zeroed in the frequency domain rather than the spatial domain. The method is designed to yield uncertainty estimates and improved generalization by stochastically suppressing frequency bands that may correspond to dataset-specific noise or bias, especially in imaging tasks such as semantic medical segmentation. Compared to standard MC Dropout—which operates directly in the signal or activation space and behaves as local impulse noise—MC Frequency Dropout (MC-FD) induces smooth global texture variations and better preserves anatomical or structural details, leading to more faithful and well-calibrated uncertainty quantification (Zeevi et al., 20 Jan 2025).

1. Motivation and Rationale

Traditional MC Dropout randomly masks activations (features) in the network’s spatial domain, resulting in localized, non-uniform perturbations akin to impulse noise. This approach can introduce artifacts and distorts edge information, particularly problematic in semantic segmentation and medical imaging where anatomical borders are crucial. Moreover, such spatial dropout does not reflect the true noise characteristics of many imaging modalities (e.g., MRI, CT, X-ray), which are fundamentally frequency-based due to the physical and computational processes underlying image acquisition.

MC Frequency Dropout is motivated by the hypothesis that stochastic suppression of random frequency components better aligns with the noise and uncertainty structure present in real-world imaging systems. This allows for the induction of plausible, globally coherent feature variation during MC sampling, producing uncertainty maps that are less noisy and more anatomically meaningful. Empirical evidence confirms that frequency-domain perturbations yield less spatial “bleeding” of uncertainty and concentrate the most uncertainty at actual object boundaries (Zeevi et al., 20 Jan 2025).

2. Mathematical Description

Let $X \in \mathbb{R}^{m \times n}$ be a feature map. A standard convolutional operation with learned weights $W$ and bias $b$ produces output $X * W + b$. Denote by $\mathcal{F}\{\cdot\}$ the (Fast) Fourier Transform and by $\mathcal{F}^{-1}\{\cdot\}$ its inverse.

Frequency Dropout is defined as:

$$C_\theta^\mathcal{F}(X) = \sigma\left( \mathcal{F}^{-1} \left( \mathcal{F}[X * W] \odot M \right ) + b \right)$$

where $M \in \{0,1\}^{m' \times n'}$ is a mask whose elements $M_{ij} \sim \mathrm{Bernoulli}(p)$ are sampled independently with dropout rate $p$, and $W$0 denotes element-wise multiplication in the frequency domain. For MC inference, $W$1 forward passes are performed, each with an independent $W$2, yielding an empirical predictive distribution

$W$3

for softmax outputs $W$4 and pixel/voxel index $W$5 (Zeevi et al., 20 Jan 2025).

A related non-MC formulation—Frequency Dropout as a regularizer—randomly applies linear filters (sampled from Gaussian, LoG, or Gabor families) at each channel and training iteration:

$W$6

with $W$7, and $W$8 random filter kernels (Islam et al., 2022).

3. Algorithmic Integration and Placement

MC Frequency Dropout is typically integrated as follows:

• Inference-time augmentation: Dropout layers are inserted after each convolutional (encoder/decoder) block, but they operate only during inference (test) time, creating MC samples without retraining or adapting weights.
• Variants: Three layer placement variants are commonly evaluated:
  1. Encoder dilution: after encoder blocks.
  2. Decoder dilution: after decoder blocks, except the final output.
  3. Global dilution: after every encoder and decoder block (except output).

• Forward pass pseudocode:

$b$6 At inference, $W$9 such passes are performed to estimate the predictive mean and uncertainty (Zeevi et al., 20 Jan 2025).

Dropout rates are typically in the range $b$0–$b$1, with lower rates for stable predictions and higher rates enhancing variability for more robust uncertainty estimation. No adaptation or retraining is required—the network weights are fixed, and only the test-time behavior changes as a result of stochastic frequency masking (Zeevi et al., 20 Jan 2025).

4. Impact on Uncertainty Estimation and Segmentation

Empirical evidence from multiple semantic segmentation tasks supports the following effects of MC Frequency Dropout (Zeevi et al., 20 Jan 2025):

• Uncertainty calibration: Frequency Dropout achieves lower Expected Uncertainty Calibration Error (UCE) compared to spatial (signal) MC Dropout, indicating improved alignment between the model's uncertainty and the location of true prediction errors, especially along semantic boundaries.
• Boundary preservation: Uncertainty maps generated with Frequency Dropout concentrate uncertainty at object boundaries and yield a reduction (≈18%) in the spatial extent of top uncertainty regions compared to spatial dropout.
• Segmentation accuracy: Dice Similarity Coefficient (DSC) divergences from the baseline are consistently smaller or comparable to standard MC Dropout. For example, in liver tumor segmentation, Frequency Dropout led to +1.23% to +3.54% improvement in DSC vs. as much as –14.53% for signal dropout at matched MC sample sizes.
• Convergence efficiency: Low UCE is achieved with fewer MC samples (as low as $b$2), whereas signal dropout may require $b$3 for comparable stability.

Representative quantitative comparisons:

Task	Baseline DSC	Signal Dropout ΔDSC	FreqDropout ΔDSC	UCE Reduction
Liver Tumors (CT)	0.606	–14.53% — –11.72%	+1.23% — +3.54%	~20%
Lungs (X-ray)	0.831	+0.44% — +0.63%	+0.91% — +1.05%	—

This suggests MC Frequency Dropout confers both superior uncertainty quantification and practical resilience in segmentation accuracy for challenging imaging modalities.

5. Theoretical and Practical Considerations

MC Frequency Dropout applies independent Bernoulli sampling to the Fourier coefficients of feature maps at each inference pass. The resulting global frequency perturbations induce correlated, physically plausible variations in spatial texture, as opposed to the unstructured impulse noise of spatial dropout. This not only preserves anatomical coherence, but aligns the predictive posterior with true model uncertainty related to domain-specific acquisition artifacts.

Practical aspects include:

• Computational cost: FFT/IFFT operations introduce $b$4 complexity, which is more demanding than native spatial dropout for large feature maps.
• Parameter choice: Optimal dropout rates ($b$5) and layer placement are dataset- and architecture-dependent; task-specific calibration is often necessary.
• No effect on training: When used for MC inference, Frequency Dropout operates purely at test time, obviating the need for retraining.
• Implementation: Replacement or supplementation of spatial dropout with Frequency Dropout in existing deep learning frameworks is supported via custom layers performing FFT-domain masking.

6. Comparative Summary of Related Frequency Dropout Approaches

A related approach, Frequency Dropout as a feature-level regularizer, utilizes classical filters (Gaussian blur, Laplacian of Gaussian, Gabor) randomly sampled and applied channel-wise to intermediate feature maps during training (Islam et al., 2022). Unlike MC Frequency Dropout, which modifies inference statistics, this regularization is primarily geared towards preventing shortcut learning and improving cross-domain generalization by discouraging over-reliance on spurious frequency artifacts.

Key empirical findings from (Islam et al., 2022):

Model & Task	Baseline Accuracy	FD–Random Filter (FD–RF)	Robustness (Corrupted Data)
ResNet-18, CIFAR-100	65.31%	68.20%	+4.4% improvement
Cardiac MRI segmentation (Dice)	65.97%	67.66%	—

The stochastic filtering regularizer can be used in distinct or complementary settings to inference-time MC Frequency Dropout.

7. Limitations and Potential Extensions

Current limitations include:

• Computation: Repeated FFT/IFFT at each MC pass incurs additional computational overhead per inference, particularly notable for high-resolution feature maps.
• Adaptation: The effectiveness of frequency dropout can be sensitive to the choice of frequency bands targeted, the structure of the underlying feature maps, and specifics of the layer placement.
• Task generality: While strong performance is observed in medical image segmentation, broader applicability requires further empirical confirmation.

Potential future directions include adaptive or learnable frequency dropout masks, multi-resolution (e.g., wavelet) frequency perturbations, dynamic rate schedules, and integration with joint training of dropout parameters (Zeevi et al., 20 Jan 2025). A plausible implication is that further tailoring of frequency-domain regularization to imaging physics or domain properties could lead to even greater robustness and fidelity of uncertainty estimation.